Improving Traffic Safety through Traffic Accident Risk Assessment

Abstract

The continuous development of sensors and the Internet of Things has produced a large amount of traffic data with location information. The improvement of traffic safety benefits from the availability of traffic accident data. Managers can patrol and control relevant areas in advance with limited police resources, according to the short-term traffic accident predictions. As a result, the possibility of accidents can be reduced, and the level of traffic safety can be improved. The traditional approach to accident prediction relies too much on the subjective experience of traffic managers. Inspired by the deep learning technology in the field of computer vision, this study first divides the road network into regular grids and takes the number of traffic accidents in each grid as the pixel value of an image. Then, a traffic accident prediction approach based on a bi-directional ConvLSTM U-Net with densely connected convolutions (BCDU-Net) is proposed. This method mines the regular information hidden in the accident data and introduces densely connected convolutions to further extract the deep spatial-temporal features contained in the traffic accident sequence. Thus, the issues of gradient disappearance and model over-fitting caused by the traditional model in model training can be avoided. Finally, the simulation experiment is carried out on the historical traffic accident data of Yinzhou District, Ningbo City. Results show that BCDU-Net has better accuracy and precision than other models in three data sets: motor vehicle accidents, non-motor vehicle accidents, and single-vehicle accidents. Therefore, the BCDU-Net is more suitable for traffic accident prediction and has good application prospects for improving road safety.